The paper makes long video generation much faster and lighter on memory by cutting out repeated work in attention.
LongVie 2 is a video world model that can generate controllable videos for 3–5 minutes while keeping the look and motion steady over time.